Function npyv_floor_f32

numpy/core/src/common/simd/sse/math.h:418–439

    #define npyv_floor_f64 _mm_floor_pd
#else
    NPY_FINLINE npyv_f32 npyv_floor_f32(npyv_f32 a)
    {
        const __m128 one = _mm_set1_ps(1.0f);
        const __m128 szero = _mm_set1_ps(-0.0f);
        const __m128i exp_mask = _mm_set1_epi32(0xff000000);

        __m128i nfinite_mask = _mm_slli_epi32(_mm_castps_si128(a), 1);
                nfinite_mask = _mm_and_si128(nfinite_mask, exp_mask);
                nfinite_mask = _mm_cmpeq_epi32(nfinite_mask, exp_mask);

        // eliminate nans/inf to avoid invalid fp errors
        __m128 x = _mm_xor_ps(a, _mm_castsi128_ps(nfinite_mask));
        __m128i roundi = _mm_cvtps_epi32(x);
        __m128 round = _mm_cvtepi32_ps(roundi);
        __m128 floor = _mm_sub_ps(round, _mm_and_ps(_mm_cmpgt_ps(round, x), one));
        // respect signed zero
        floor = _mm_or_ps(floor, _mm_and_ps(a, szero));
        // if overflow return a
        __m128i overflow_mask = _mm_cmpeq_epi32(roundi, _mm_castps_si128(szero));
        // a if a overflow or nonfinite
        return npyv_select_f32(_mm_or_si128(nfinite_mask, overflow_mask), a, floor);
    }
    NPY_FINLINE npyv_f64 npyv_floor_f64(npyv_f64 a)
    {
        const __m128d one = _mm_set1_pd(1.0f);